A deep learning underwater image enhancement method based on dual-path joint correction

Abstract

The present invention provides a deep learning underwater image enhancement method based on dual-path joint correction, comprising the following steps: obtaining a basic feature map; obtaining feature maps of R, G, and B channels respectively; adding a channel attention mechanism to obtain a weight matrix and multiplying it to each pixel value of the corresponding channel on the previous feature map; obtaining representative color features; performing two-dimensional attention processing on the basic feature map obtained in the texture feature extraction path; forming a recalibrated texture detail feature for the high-order feature through a pixel attention module; splicing the color feature and the texture feature into a new comprehensive feature, and feeding it into a multi-scale U-Net network; and finally outputting the enhanced result image. The present invention adopts a triple feature extraction module to process the R, G, and B channels of the image respectively, balances the grayscale ratio of the R, G, and B colors of the degraded image, solves the color cast problem of the image, and obtains a better color correction effect.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a deep learning underwater image enhancement method based on dual-path joint correction. Background Art

[0002] In recent years, with the scarcity of land resources, the precious resources contained in the ocean world have gradually become the focus, and people have begun to actively formulate ocean strategies. Underwater images are an important medium for marine information. The quality of underwater images is of great significance for the exploration and utilization of the deep sea, such as target detection, seabed minerals and marine energy exploration. However, underwater imaging is affected by complex underwater scenes and lighting conditions, such as the exponential growth of underwater light attenuation and different light attenuation rates. Underwater images generally have problems such as blurred texture, low contrast, and color distortion. Many researchers are committed to image enhancement methods to obtain high-quality underwater images to explore the underwater world.

[0003] Due to the complexity of the underwater environment, degraded underwater images cannot accurately reflect the underwater scene information. Due to the complexity of the underwater scene and its accompanying light sources, the absorption of light by water and the scattering of light by suspended particles are the main reasons for the attenuation of light energy. Due to the different attenuation rates of light of different wavelengths, red light attenuates the fastest when propagating in water, while blue-green light is the slowest attenuating light, so the collected underwater images present a blue-green hue. The attenuated direct transmission causes the scene intensity to weaken, and the surrounding scattered light causes the image's real scene color distortion, low contrast, and blurred details. In order to solve the above problems, many recent works have proposed various effective underwater image enhancement and restoration methods from different perspectives.

[0004] Traditional methods based on non-physical models do not consider the relevant complex factors of the underwater imaging model and directly improve the visual effect of the image by adjusting the pixel values ​​of the image. Traditional methods based on physical models rely on the underwater imaging model to generate clear images by estimating atmospheric light and transmission maps. Learning-based methods accurately extract feature information to restore degraded images, overcoming the shortcomings of traditional methods, such as insufficient robustness of detail information recovery. Learning-based methods are combined with physical models, using convolutional neural networks to obtain transmission maps and atmospheric light, improving the accuracy of estimated parameters, and then performing a reversible process to obtain the restored underwater image.

[0005] Although traditional underwater image enhancement methods ignore the optical characteristics of underwater imaging, they are prone to introduce color casts, artifacts, and aggravated noise. Traditional underwater image restoration methods rely on specific priors, which leads to the complexity and limitations of estimated parameters and cannot effectively solve problems such as image color distortion and artificial light source removal.

[0006] According to research, various underwater imaging technologies and methods have been introduced into the field of underwater image processing. When traditional methods are ineffective for some obviously distorted images or the processing effect is not ideal due to technical limitations, deep learning technology has become the mainstream method for image enhancement. The emergence of deep learning has greatly improved low-level visual tasks. The end-to-end network structure is suitable for most multi-classification scenarios, and can achieve good processing effects for most scenarios. Therefore, deep learning underwater enhancement technology has been widely used. Summary of the invention

[0007] According to the technical problems mentioned in the above-mentioned background technology, a deep learning underwater image enhancement method based on dual-path joint correction is provided. The present invention mainly processes the original underwater image through the R, G, and B channels separately through a ternary color feature extraction module to obtain the color features of different channels for color correction. The joint channel attention module highlights important features while suppressing less important features to recalibrate color features. At the same time, the designed two-dimensional attention module is used to extract texture features with high precision and refine background information to restore the texture details of the image. Finally, the image is enhanced through a multi-scale U-Net adaptive structure.

[0008] The technical means adopted by the present invention are as follows:

[0009] A deep learning underwater image enhancement method based on dual-path joint correction includes the following steps:

[0010] Step S01: Feed the input original image to the color feature extraction path and the texture feature extraction path respectively, perform 3*3 convolution operations on the two paths respectively, and obtain the basic feature map;

[0011] Step S02: Decomposing the basic feature map according to R, G, and B channels respectively to obtain R, G, and B channel data of the image, and then learning the R, G, and B channel data of the image separately through three primary color channels to obtain feature maps of the R, G, and B channels respectively;

[0012] Step S03: For the feature maps of the R, G, and B channels, a channel attention mechanism is added, pixel values ​​are compressed by global average pooling, and then a real number of 0 to 1 is normalized by a Sigmoid activation function, where 1 corresponds to importance and 0 corresponds to unimportance, and the obtained real number is multiplied by each pixel value of the corresponding channel on the previous feature map; and the weights assigned to the features of each channel are adaptively adjusted according to the above;

[0013] Step S04: performing a two-dimensional attention process on the basic feature map obtained in the texture feature extraction path; firstly, the basic feature map is passed through a visual feature attention module, and then a double pooling operation is used to further extract high-order features to visually enrich feature details;

[0014] Step S05: extracting internal information from the high-order features through a pixel attention module to form a recalibrated texture detail feature;

[0015] Step S06: The color features and texture features obtained in step S03 and step S04 are spliced ​​into a new comprehensive feature, and fed into a multi-scale U-Net network;

[0016] Step S07: In the U-Net encoding stage, the comprehensive features described in step S06 are downsampled to obtain a latent vector. In the U-Net decoding stage, the latent vector is upsampled to the original size of the image. The image after each upsampling is jump-connected with the output of the symmetric layer of the corresponding encoding stage, and finally the enhanced result image is output.

[0017] Step 08: Finally, use the l1 loss function to improve the performance of the overall network.

[0018] Compared with the prior art, the present invention has the following advantages:

[0019] The present invention adopts a triplet feature extraction module to process the R, G, and B channels of the image separately, balances the grayscale ratio of the R, G, and B colors of the degraded image, solves the color cast problem of the image, and obtains a better color correction effect.

[0020] The present invention uses a visual feature attention module and a double pooling operation to aggregate feature information, thereby reducing the error of increased estimate variance due to the limited neighborhood size during feature extraction and the mean shift error caused by the parameter error of the convolutional layer during the feature extraction process, aiming to retain rich texture details so that the final enhanced image is more in line with human subjective visual effects.

[0021] Based on the above reasons, the present invention can be widely promoted in the fields of image processing and the like. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0023] Figure 1It is a schematic diagram of the process of the present invention.

[0024] Figure 2 This is a comparison of the enhancement effects of the present invention and other algorithms on underwater scene images. Figure 2-1 -1. Figure 2-2 -1. Figure 2 -3-1, Figure 2 -4-1 is the result image after being processed by the UWGAN algorithm; Figure 2-1 -2. Figure 2-2 -2. Figure 2 -3-2, Figure 2 -4-2 is the result image after being processed by the FUnIE-GAN algorithm; Figure 2-1 -3. Figure 2-2 -3. Figure 2 -3-3, Figure 2 -4-3 is the result image after being processed by Ucolor algorithm; Figure 2-1 -4. Figure 2-2 -4. Figure 2 -3-4, Figure 2 -4-4 is the result image after being processed by the UWCNN algorithm; Figure 2-1 -5. Figure 2-2 -5. Figure 2 -3-5, Figure 2 -4-5 is the result diagram after being processed by the algorithm proposed in this article. DETAILED DESCRIPTION

[0025] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0026] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0027] like Figure 1 As shown, the present invention provides a deep learning underwater image enhancement method based on dual-path joint correction, comprising the following steps:

[0028] Step S01: Feed the input original image to the color feature extraction path and the texture feature extraction path respectively, perform 3*3 convolution operations on the two paths respectively, and obtain the basic feature map;

[0029] Step S02: Decompose the basic feature map according to the R, G, and B channels respectively, obtain the R, G, and B channel data of the image, and then learn the R, G, and B channel data of the acquired image separately through the three-primary color channel to obtain the feature maps of the R, G, and B channels respectively. In the present application, the three-primary color feature extraction module balances the grayscale ratio of RGB of the degraded image and adjusts the distribution of RGB, including multiple convolution layers with a kernel of 3*3 and splicing operations;

[0030]

[0031]

[0032]

[0033] CF=Cat(CF r ,CF g ,CF b )∈R N×H×W ; (4);

[0034] Among them, * represents the convolution operation, Cat represents the concatenation operation, H and W represent the height and width of the feature map, N represents the number of feature maps, and CF represents the color feature.

[0035] Step S03: For the feature maps of the R, G, and B channels, add a channel attention mechanism, compress the pixel values ​​through global average pooling, and then normalize the real numbers from 0 to 1 through the Sigmoid activation function in the activation operation, where 1 corresponds to important and 0 corresponds to unimportant, and multiply the obtained real number to each pixel value of the corresponding channel on the previous feature map; and adjust the weights assigned to the features of each channel according to the above adaptive adjustment;

[0036] Step S04: Perform two-dimensional attention processing on the basic feature map obtained in the texture feature extraction path; first pass the basic feature map through the visual feature attention module, and then use the double pooling operation to further extract high-order features to visually enrich the feature details. Preferably, in the process of combining the channel attention module, first use global average pooling to compress the features to obtain the feature map of the global receptive field, and stretch the size of the feature map to 1×1×C. This process is defined as:

[0037]

[0038] Among them, D k represents the feature descriptor, F sq Indicates the compression Squeeze operation, CF k Represents color features, H and W represent the height and width of the feature map respectively;

[0039] First, a fully connected operation is used to perform a nonlinear transformation on the compression result in the activation step; then, the parameters W1 and W2 are learned and weights are generated for each feature channel; the new feature U obtained after the activation operation is:

[0040] U=F ex (D,W)=σ(h(D,W))=σ(W2*δ(W1*D)); (6)

[0041] Among them, F ex represents the activation operation, δ represents the ReLU activation function, σ represents the Sigmoid activation function, * represents the convolution operation, W1 and W2 represent the weights, and U represents the reconstructed features after the activation operation; the final output of the channel attention module is reconstructed as:

[0042] CF=F scale (U, CF) = CF·U (7);

[0043] Among them, F scale (U,CF) represents the feature map CF∈R N×H×W The channels between y and U are multiplied to complete the recalibration of the original features in the channel dimension.

[0044] The overall process of extracting texture features by the two-dimensional attention module is defined as:

[0045]

[0046]

[0047] Among them, MAP v Represents the visual attention map, MAP srepresents the pixel attention map; BF represents the basic features extracted by the filter, Represents the product of corresponding elements; F1 represents the features reconstructed by the visual feature attention module, and F2 represents the final reconstructed output feature vector.

[0048] The visual feature attention module includes: global average pooling, global maximum pooling, splicing operation and activation function; the global average pooling and global maximum pooling are used to generate different feature descriptors, respectively represented as maximum pooling layers and average pooling layer Then, the elements are summed up to output a new feature, and finally, the sigmoid excitation function is used to improve the performance of the local network. In short, this process can be described as:

[0049]

[0050] Among them, MAP v represents the visual attention map, σ represents the Sigmoid activation function, Maxpool and Avgpool represent the maximum pooling operation and the average pooling operation, f 3×3 represents a convolution operation with a kernel size of 3*3, BF is the basic feature, and Represents the maximum pooling layer and the average pooling layer.

[0051] Step S05: The pixel attention module extracts internal information from high-order features to form recalibrated texture detail features; the pixel attention module applies global maximum pooling and global average pooling in the spatial dimension, and then connects the pooled features to generate new effective feature descriptors; 2D convolution is used to aggregate the information on the connected feature descriptors to generate a pixel attention map that can encode the focus or suppression MAP; the calculation process of MAPs is:

[0052]

[0053] Where σ represents the Sigmoid activation function and * represents the convolution operation with a kernel size of 3*3.

[0054] Step S06: The color features and texture features obtained in step S03 and step S04 are spliced ​​into a new comprehensive feature, and fed into a multi-scale U-Net network;

[0055] Step S07: In the U-Net encoding stage, the comprehensive features of step S06 are downsampled to obtain a latent vector. In the U-Net decoding stage, the latent vector is upsampled to the original size of the image. Each upsampled image is jump-connected to the output of the symmetric layer of the corresponding encoding stage, and finally the enhanced result image is output.

[0056] Step 08: Finally, the l1 loss function is used to improve the performance of the overall network; the l1 loss function represents the minimum absolute deviation and measures the sum of all absolute differences between the true value and the predicted value; the l1 loss function is specifically defined as:

[0057]

[0058] Where x represents the index of the pixel in region X, represents the reconstructed image pixel, and J(x) represents the corresponding real image pixel value.

[0059] Example 1

[0060] like Figure 2 As shown, the first image is first processed by a variety of algorithms; Figure 2-1 -1 is the result image after being processed by the UWGAN algorithm. It can be seen that the result image has a serious foggy feeling, and the blurred details lead to extremely low visibility, which does not achieve the expected effect. Figure 2-1 -2 is the result image after being processed by the FUnIE-GAN algorithm. It can be seen that the result image removes the fog, but the image texture details are unnatural and blurred, the overall contrast is low, and the effect needs to be improved. Figure 2-1 -3 is the result image after being processed by the Water-Net algorithm. It can be seen that the blurring of texture details in the result image has been resolved and the contrast has been partially improved, but the overall brightness of the image is dark and the expected effect has not been achieved. Figure 2-1 -4 is the result image after being processed by the UWCNN algorithm. It can be seen that the result image has serious color cast problems, poor contrast and object restoration, and the processing result is very poor. Figure 2-1 -5 is the result image after being processed by the algorithm in this paper. It can be found that the color cast problem has been solved, the texture details are restored to a high degree, the contrast is effectively enhanced, and the image enhancement is relatively successful.

[0061] Then process the second image. Figure 2-2 -1 is the result image after being processed by the UWGAN algorithm. It can be seen that the details of the result image are lost, the contrast is extremely low, and the effect is not good. Figure 2-2 -2 is the result image after being processed by the FUnIE-GAN algorithm. It can be seen that the result image is still gray as a whole, the image still has unclear texture, and the color saturation and contrast do not meet the expected effects. Figure 2-2-3 is the result image after being processed by the Water-Net algorithm. It can be seen that the color cast problem of the result image has been partially improved, and the contrast has been partially improved, but the visibility of the objects in the image is low, and the foreground objects are difficult to identify, which does not achieve the expected effect. Figure 2-2 -4 is the result image after being processed by the UWCNN algorithm. It can be seen that the brightness of the result image is seriously dim, the object distortion is high, and the processing result is difficult to be satisfactory. Figure 2-2 -5 is the result image after being processed by the algorithm in this paper. It can be found that the color cast problem has been solved, the contrast has been effectively enhanced, the image restoration is natural and realistic, and the image enhancement is relatively successful.

[0062] Then process the third image. Figure 2 -3-1 is the result image after being processed by the UWGAN algorithm. It can be seen that the color cast is still serious, the background details are very blurred, the contrast is extremely low, and the effect is not good. Figure 2 -3-2 is the result image after being processed by the FUnIE-GAN algorithm. It can be seen that the result image is grayish overall, and the contrast is significantly improved, but there are slight artifacts and the expected effect is not achieved. Figure 2 -3-3 is the result image after being processed by the Water-Net algorithm. It can be seen that the artifact problem of the result image has been partially improved, and the details and contrast have been partially improved, but the brightness is too low and does not conform to human visual effects. Figure 2 -3-4 is the result image after being processed by the UWCNN algorithm. It can be seen that the result image has a serious color cast problem, the detail information is basically impossible to obtain, the degree of restoration of the object is extremely low, and the processing result is very poor. Figure 2 -3-5 is the result image after being processed by the algorithm in this paper. It can be found that the color cast problem has been solved, the details have been significantly improved, the contrast has been effectively enhanced, the visibility of the object has been improved, the artifacts have been removed, and the image enhancement is relatively successful.

[0063] Then process the fourth image. Figure 2 -4-1 is the result image after being processed by the UWGAN algorithm. It can be seen that the problem of blurred details in the result image is still serious, the contrast is extremely low, and the processing effect is poor. Figure 2 -4-2 is the result image after being processed by the FUnIE-GAN algorithm. It can be seen that the result image is still dark. Although the contrast has been improved, it is still low and has not achieved the expected effect. Figure 2 -4-3 is the result image after being processed by the Water-Net algorithm. It can be seen that the color cast problem of the result image has been partially improved, but there is still a problem of overall low brightness, blurred information at the edges, and unable to obtain more detailed information, and the expected effect has not been achieved. Figure 2-4-4 is the result image after being processed by the UWCNN algorithm. It can be seen that the result image has a serious color cast problem, detailed information cannot be obtained, and the processing result is difficult to be satisfactory. Figure 2 -4-5 are the result images after being processed by the algorithm in this paper. It can be found that the color cast problem has been solved, the details have been significantly improved, the contrast has been effectively enhanced, and the image enhancement is relatively successful. It can be observed that the first four algorithms are insufficient in color correction, contrast and details, while the algorithm proposed in this paper has effectively processed the color cast, low contrast and blurred details of underwater images, corrected the image color, improved the contrast and highlighted the details; it shows that the degraded image has been well enhanced.

[0064] In order to ensure the reliability of the evaluation, this embodiment conducts comparative analysis from four objective indicators: PSNR, SSIM, UCIQE and UIQM. For specific data, see Table 1. It can be seen from the indicators that the method of the present invention is superior to the objective evaluation indicators of other methods, proving that the method of the present invention can effectively improve the texture details, color richness and contrast of the image.

[0065] Table 1 Comparison of the processing results of the algorithm of the present invention and other algorithms

[0066] method PSNR SSIM UCIQE UIQM UWGAN 14.235 0.695 0.499 0.989 FUnIE-GAN 16.399 0.585 0.517 1.063 Water-Net 19.165 0.817 0.571 1.337 UWCNN 13.054 0.616 0.455 1.083 Ours 21.817 0.879 0.590 1.353

[0067] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0068] In the above embodiments of the present invention, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0069] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A deep learning underwater image enhancement method based on dual-path joint correction, characterized in that: The following steps are involved: Step S01: Feed the input original image to the color feature extraction path and the texture feature extraction path respectively, perform 3*3 convolution operations on the two paths respectively, and obtain the basic feature map; Step S02: Decomposing the basic feature map according to R, G, and B channels respectively to obtain R, G, and B channel data of the image, and then learning the R, G, and B channel data of the image separately through three primary color channels to obtain feature maps of the R, G, and B channels respectively; Step S03: For the feature maps of the R, G, and B channels, a channel attention mechanism is added, pixel values ​​are compressed by global average pooling, and then a real number of 0 to 1 is normalized by a Sigmoid activation function, where 1 corresponds to importance and 0 corresponds to unimportance. The obtained real number is multiplied by each pixel value of the corresponding channel on the previous feature map, and the weights assigned to the features of each channel are adjusted; Step S04: performing a two-dimensional attention process on the basic feature map obtained in the texture feature extraction path; firstly, the basic feature map is passed through a visual feature attention module, and then a double pooling operation is used to further extract high-order features to visually enrich feature details; Step S05: extracting internal information from the high-order features through a pixel attention module to form a recalibrated texture detail feature; Step S06: The color features and texture features obtained in step S03 and step S04 are spliced ​​into a new comprehensive feature, and fed into a multi-scale U-Net network; Step S07: In the U-Net encoding stage, the comprehensive features described in step S06 are downsampled to obtain a latent vector. In the U-Net decoding stage, the latent vector is upsampled to the original size of the image. Each upsampled image is jump-connected to the output of the symmetric layer of the corresponding encoding stage, and finally an enhanced result image is output; Step S08: Finally, the l1 loss function is used to improve the performance of the overall network.

2. The method for underwater image enhancement based on deep learning and dual-path joint correction according to claim 1, characterized in that: In the step S01, the grayscale ratio of RGB of the degraded image is balanced and the distribution of RGB is adjusted, including multiple convolution layers with a kernel of 3*3 and a splicing operation; CF=Cat(CF r ,CF g ,CF b )∈R N×H×W ;(4); Among them, * represents the convolution operation, Cat represents the concatenation operation, H and W represent the height and width of the feature map, N represents the number of feature maps, and CF represents the color feature.

3. The method for underwater image enhancement based on deep learning and dual-path joint correction according to claim 1, characterized in that: The step S03 comprises the following steps: First, use global average pooling to compress the features to obtain the feature map of the global receptive field, and stretch the size of the feature map to 1×1×C. This process is defined as: Among them, D k represents the feature descriptor, F sq Indicates the compression Squeeze operation, CF k Represents color features, H and W represent the height and width of the feature map respectively; First, a fully connected operation is used to perform a nonlinear transformation on the compression result in the activation step; then, the parameters W1 and W2 are learned and weights are generated for each feature channel; the new feature U obtained after the activation operation is: U=F ex (D,W)=σ(h(D,W))=σ(W2*δ(W1*D));(6) Among them, F ex represents the activation operation, δ represents the ReLU activation function, σ represents the Sigmoid activation function, * represents the convolution operation, W1 and W2 represent the weights, and U represents the reconstructed features after the activation operation; the final output of the channel attention module is reconstructed as: Among them, F scale (U,CF) represents the feature map CF∈R N×H×W The channels between y and U are multiplied to complete the recalibration of the original features in the channel dimension.

4. The method for underwater image enhancement based on deep learning and dual-path joint correction according to claim 1, characterized in that: In step S04, the overall process of performing two-dimensional attention processing to extract texture features is defined as: Among them, MAP v Represents the visual attention map, MAP s represents the pixel attention map; BF represents the basic features extracted by the filter, Represents the product of corresponding elements; F1 represents the features reconstructed by the visual feature attention module, and F2 represents the final reconstructed output feature vector.

5. The method for underwater image enhancement based on deep learning and dual-path joint correction according to claim 1, characterized in that: The visual feature attention module includes: global average pooling, global maximum pooling, splicing operation and activation function; the global average pooling and global maximum pooling are used to generate different feature descriptors, respectively represented as maximum pooling layers and average pooling layer Then, the elements are summed up to output a new feature, and finally, the sigmoid excitation function is used to improve the performance of the local network; this process can be described as: Among them, MAP v represents the visual attention map, σ represents the Sigmoid activation function, Maxpool and Avgpool represent the maximum pooling operation and the average pooling operation, f 3×3 represents a convolution operation with a kernel size of 3*3, BF is the basic feature, and Represents the maximum pooling layer and the average pooling layer.

6. The method for underwater image enhancement based on deep learning and dual-path joint correction according to claim 1, characterized in that: The pixel attention module applies global maximum pooling and global average pooling in the spatial dimension, and then connects the pooled features to generate new valid feature descriptors; 2D convolution is used to aggregate the information on the connected feature descriptors to generate pixel attention maps that can encode the focus or suppression MAPs; the calculation process of MAPs is: Where σ represents the Sigmoid activation function and * represents the convolution operation with a kernel size of 3*3.

7. The method for underwater image enhancement based on deep learning and dual-path joint correction according to claim 1, characterized in that: The loss function is an l1 loss function; the l1 loss function represents the minimum absolute deviation and measures the sum of all absolute differences between the true value and the predicted value; the l1 loss function is specifically defined as: Where x represents the index of the pixel in region X, represents the reconstructed image pixel, and J(x) represents the corresponding real image pixel value.

Images